Invisibility Cloak for object detection

This research dated back to 2020. The researchers from University of Maryland Computer Science published their studies on creating adversarial attacks on object detectors, including a sweater whose printed pattern hides its wearer from a YOLOv2 person detector. The research paper is available under 10.48550/arXiv.1910.14667. It was accepted to ECCV 2020 and tests the patterns against detector ensembles in white-box and black-box settings, on printed posters as well as on wearable clothing. Hackaday picked up the “adversarial sweater of doom” and points out that the goal is probing machine learning models for weaknesses, not selling you a practical cloak. Fashion-wise, it is an ugly sweater either way.